From: Robert Brunner (rbrunner_at_illinois.edu)
Date: Fri Sep 11 2009 - 11:34:12 CDT
The issue is the communication:computation ratio. With lower-latency
communication (e.g. infiniband instead of ethernet), you can utilize
more processors efficiently with fewer atoms/processor. With faster
processors, you need more atoms/processor, since the computation takes
500-1000 atoms is a rule of thumb, not a law, so for any particular
simulation and computer configuration, its always good to run a few
tests with various numbers of processors and figure out how many
processors can be used effectively. Make sure, however, that any
performance comparisons are based on timings after NAMD has done the
first few load balancing cycles.
Note that NAMD efficiency is much more sensitive to communication
latency, rather than bandwidth. This can be a factor when you're
building a new cluster, and are deciding how much to spend on
communication hardware vs. CPU power or memory.
If you're saving restart files or DCD frames every minute or so (or
less frequently), it shouldn't be an issue. If you're saving data
every step, then it probably slows things down.
On Sep 11, 2009, at 8:22 AM, Amr Zeinalabideen Majul wrote:
> Hello all,
> Im simulating a small system of 16000 atoms, on a different number
> of CPUS to get an idea of the performance increases. I noticed that
> at a 100 CPUs, that only 10-20 percent of the CPU is being used.
> I searched around and I found out that estimates that each cpu can
> handle a minimum of around 500 to 1000 atoms, I was wondering if
> this is a reliable method of roughly estimating how many CPUs I
> need, and if this is indeed the reason why my CPUs are not working
> at maximum efficiency? If indeed each CPU can handle 500 atoms with
> linear increases it would seem more than 32 cpus isnt much help.
> Does that sound correct?
> Or would it be a communication issue between the cpus? How much
> would the frequency of output to the various files affect the
> wallclock time of the simulation if I am using ethernet to
> communicate between the nodes?
> Amr Majul
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:53:16 CST